Hi. I read two memory update methods in your two version papers. It looks like updating with the hardest sample gives better performance, but the selection of hyperparameters is tricky. The latest version of the paper proposed the cm method considering all query images, which may be more robust. Do you think it is possible to mix these two methods like applying them randomly during training or at the first n epochs uses a certain method and the rest of epochs uses another method? Thanks.
Hi. I read two memory update methods in your two version papers. It looks like updating with the hardest sample gives better performance, but the selection of hyperparameters is tricky. The latest version of the paper proposed the cm method considering all query images, which may be more robust. Do you think it is possible to mix these two methods like applying them randomly during training or at the first n epochs uses a certain method and the rest of epochs uses another method? Thanks.